Breaking
RUAmerican actress Catherine Cowan attended the amfAR charity evening in Venice in a revealing outfitRURussian aviation destroyed the underground base of the Ukrainian Armed Forces in Kherson with a FAB-1500 bombRUIn Saratov, as a result of a UAV attack, there was damage to civilian infrastructure and casualtiesRUWhite House names Secretary of State Marco Rubio Labor Day Employee of the MonthRUCaptured Ukrainian soldier reveals payments for killing Russian servicemenCN"Ask" about health丨Autumn dryness is coming. How to replenish water correctly for the body?CNAfter the storm, hope is rising (On the spot·Special report)CNTaiwan's Foundation Member Wu Hsintai criticizes Li Baiyi of the Kuomintang for exercising privileges in the emergency roomCNThe 15th China Intellectual Property Annual Conference was held in Beijing. Multiple data demonstrated the results of high-quality development of intellectual property.CNZiyang 431 Community: Exploring new paths for grassroots governance through national unity, co-construction and sharingRUAmerican actress Catherine Cowan attended the amfAR charity evening in Venice in a revealing outfitRURussian aviation destroyed the underground base of the Ukrainian Armed Forces in Kherson with a FAB-1500 bombRUIn Saratov, as a result of a UAV attack, there was damage to civilian infrastructure and casualtiesRUWhite House names Secretary of State Marco Rubio Labor Day Employee of the MonthRUCaptured Ukrainian soldier reveals payments for killing Russian servicemenCN"Ask" about health丨Autumn dryness is coming. How to replenish water correctly for the body?CNAfter the storm, hope is rising (On the spot·Special report)CNTaiwan's Foundation Member Wu Hsintai criticizes Li Baiyi of the Kuomintang for exercising privileges in the emergency roomCNThe 15th China Intellectual Property Annual Conference was held in Beijing. Multiple data demonstrated the results of high-quality development of intellectual property.CNZiyang 431 Community: Exploring new paths for grassroots governance through national unity, co-construction and sharing
Back算力涨价潮影响与应对:一场席卷产业链的供需博弈
算力涨价潮影响与应对:一场席卷产业链的供需博弈
Tech
中国新闻网1 hour agoTech7 min readChinaView translation

算力涨价潮影响与应对:一场席卷产业链的供需博弈

Quick Look

今年以来,腾讯云、阿里云等云计算厂商密集上调算力产品价格。受AI算力供需失衡、核心硬件产能紧张等因素影响,算力成本上涨正向上下游传导并引发产业链加速调整。

AI-generated summary

Why It Matters

截至2024年初,我国日均Token调用量为1000亿,到2026年3月已超过140万亿。2025年我国推理数据量达101.34EB,首次超过训练数据量。

Font size

一场席卷产业链的供需博弈

——算力涨价潮影响与应对①

今年以来,腾讯云、阿里云、百度智能云等云计算厂商及人工智能(AI)服务商密集上调算力产品价格,部分服务最高涨幅超过30%。此轮算力涨价,不仅是企业的价格调整,更是一场席卷产业链的供需博弈。

算力是数字经济的核心生产要素,这一轮算力涨价潮因何而来?价格上涨如何传导至产业链上下游?又将给正在加速发展的AI产业带来怎样的影响?

三重因素推高价格

“算力价格上涨,根本原因是供需失衡。需求端快速增长,而供给端受多重瓶颈制约,两者之间的缺口持续扩大。”赛智产业研究院智能经济研究所所长李铭岩说。

国家发展改革委宏观经济研究院研究员张林山认为,生成式AI应用的井喷式发展催生了海量算力需求,科技巨头与初创企业纷纷投身“百模大战”,导致对高性能图形处理器(GPU)、智算中心的需求呈指数级激增。

词元(Token)调用量的快速增长是需求端扩张的直接体现。截至今年3月,我国日均Token调用量已超过140万亿,相比2024年初的1000亿增长了1000多倍,相比2025年底的100万亿,3个月时间增长超40%。

需求结构的转变,折射出算力市场的新趋势。国家数据局数据显示,2025年,我国用于人工智能训练和推理的数据总量为199.48艾字节(EB),推理数据量首超训练数据量,达101.34EB。据测算,未来推理算力需求与训练算力需求之比或将达到3∶1,甚至更高。训练决定模型的能力上限,而推理则决定模型的商业落地,这种转变,使算力从阶段性投入变成持续性运营支出。

供给端的约束也十分突出。国研新经济研究院创始院长朱克力认为,高端AI芯片、高带宽存储等核心硬件产能紧张,上游元器件成本抬升,直接推高服务器采购成本。同时智算中心建设还要匹配电力、网络配套,新建算力集群投产周期长,短期供给跟不上需求扩张的节奏。

李铭岩判断,此轮算力涨价不是短期波动,而是AI产业从“补贴扩张”走向“商业可持续”的标志性节点。需求端从训练向推理的结构性迁移打开了长期增长空间,供给端的硬件、封装、交付等瓶颈在短期内难以根本缓解,而行业自身的商业逻辑已从“烧钱”转向“盈利”。三重力量叠加,共同推高了算力价格。

产业链正加速调整

8月17日,DeepSeek API(应用程序编程接口)调价正式生效。业内人士分析,算力成本压力是人工智能大模型涨价的主要原因。

算力价格上涨,正沿着产业链向上下游传导。一场围绕成本、利润和市场份额的博弈也随之展开。涨价首先向上游传导,带动算力基础设施投资扩张。随着AI需求持续增长,云厂商不断加大资本开支,推动GPU、存储、先进封装等核心硬件需求扩张。“上游硬件厂商迎来量价齐升的景气周期,芯片、存储、先进封装等上游环节是涨价的最大受益者。”李铭岩说。

处于产业链中游的算力租赁企业,则面临机遇与压力并存的局面。业内人士表示,随着算力需求持续增长,算力租赁市场空间进一步扩大,相关企业的议价能力有望提升。与此同时,算力成本占其运营支出的七八成,上游价格上涨将直接压缩利润空间。

当涨价压力最终传导至下游,AI应用企业的成本承压将更加明显,行业分化或将进一步加剧。算力成本正成为大模型公司和AI应用企业的运营负担,头部模型厂商成为云服务商之外算力租赁市场的重要增量客户,但实力较弱的中小AI企业可能因算力成本过高而被挤出市场,加速行业洗牌。

张林山认为,面对算力涨价挑战,企业应采取降本增效与多元布局的策略。技术上,通过模型蒸馏、量化等手段降低算力消耗,并探索国产芯片等替代方案;资源上,采用混合云架构,灵活调度算力资源以平滑价格波动;战略上,大型企业可适度布局自建算力设施,中小企业则可利用算力租赁模式降低初始投入,通过精细化管理最大化单位算力的价值。

朱克力表示,算力供给企业,不能简单依靠涨价获利,要提升资源调度效率,盘活闲置算力,丰富按量计费、弹性租赁等多元服务模式。长远来看,整个行业要跳出单纯拼硬件资源的思路,向算法优化、场景解决方案延伸,依靠技术效率对冲硬件成本上涨带来的压力。

完善算力供给体系

算力涨价潮会持续多久?多位专家判断,未来两三年算力价格将呈现先稳中有升,而后逐步趋稳分化的走势。短期高端训练算力仍将维持高位,随着全球算力产能持续投放、国产算力生态成熟,通用推理算力涨幅放缓;长期来看,价格将逐步回落。

面对供给短缺与需求增加双重压力,除了扩大硬件供给,通过云端服务的调度能力,把已有和新增的算力资源利用到最优,成为当前许多企业的选择。

以腾讯云为例,通过统一调度层,腾讯云将不同芯片、不同任务、不同位置的算力资源纳入同一平台管理,按业务实际需要动态匹配——训练与推理弹性分配、云端与边端协同承担、离线与在线错峰运行。

中国信息通信研究院云计算与数字化研究所副所长栗蔚认为,行业将从比拼浮点运算能力,转向比拼Token效能。云服务商需通过模型架构优化等手段系统性降低Token成本,从而在价格层面构建差异化竞争优势,实现技术降本与商业竞争力的双重突破。

今年4月,算力网被正式纳入国家“六张网”建设,围绕国家算力枢纽节点已建成超70条算力大通道。李铭岩认为,应在此基础上加快推动算力资源全国“一张网”调度,打破区域和主体壁垒,让算力真正像水电一样流动自如;持续探索“算力换绿电”等市场化合作模式,让西部绿电真正成为东部算力的低成本能源支撑;加大研发投入,突破先进制程、先进封装等关键技术瓶颈。

构建高效、普惠、安全的算力供给体系需要多方协同推进。张林山建议,技术层面,持续突破高端芯片、基础软件关键技术,完善异构算力调度平台,提升算力资源利用率;布局全国一体化算力网络,深化“东数西算”工程,推动算力跨区域流通交易,降低算力使用门槛,促进算力普惠。产业层面,鼓励绿色智算中心建设,发展液冷、低碳算力基础设施。

(经济日报 记者 赖奇春 黄鑫)

What to Watch

AI outlook — possibilities, not facts

  • 未来两三年算力价格将呈先稳中有升而后逐步趋稳分化走势。

    Likely · Within months

Open Questions

  • 算力网全国“一张网”调度何时能够彻底打破跨区域和主体壁垒?
  • 国产芯片等替代方案能否在短期内填补高端AI芯片产能缺口?

Related Topics

This article was originally published by 中国新闻网.

Related Stories

The Ministry of Industry and Information Technology released the "15th Five-Year Plan for the Development of the Information and Communications Industry" to deploy 6G technology research and development and improve the resilience of the industrial chain.
Developing·

The Ministry of Industry and Information Technology released the "15th Five-Year Plan for the Development of the Information and Communications Industry" to deploy 6G technology research and development and improve the resilience of the industrial chain.

The Ministry of Industry and Information Technology issued the "15th Five-Year Plan for the Development of the Information and Communications Industry" on September 7, proposing to deploy the world's leading network facilities, launch 6G commercialization in a timely manner, accelerate 6G core technology research and development, conduct 6G technology trials, and develop 6G base stations, core network equipment and 6G smartphones to enhance the resilience of the industrial chain and supply chain.

中国新闻网
2 min read
Grindr agrees to settle UK user privacy lawsuit for £26m
Developing·

Grindr agrees to settle UK user privacy lawsuit for £26m

Grindr, the world's largest LGBTQ+ dating app, announced that it has agreed to settle a British class action lawsuit for £26 million (approximately NT$1.1 billion) over allegations that it was suspected of leaking users' sensitive personal information by the Chinese company Kunlun Vanguard before 2020. If distributed equally, each of the approximately 12,000 users would receive approximately £2,167. Although Grindr disagrees with the accusation, it said it understands the pain of users and will pay compensation in two installments.

自由时报
2 min read
On the eve of the mass production of dexterous hands, why are products with a high degree of freedom "popular but not popular"? Industry: The industry is still looking for the optimal solution
Developing·

On the eve of the mass production of dexterous hands, why are products with a high degree of freedom "popular but not popular"? Industry: The industry is still looking for the optimal solution

Although the dexterous hand industry is highly popular and frequently funded, high-degree-of-freedom products are difficult to achieve large-scale mass production due to factors such as high costs, lagging model capabilities, and dispersed demand. The market is still dominated by low-degree-of-freedom products such as three-finger claws. The industry is shifting from a hardware competition to a software-constrained stage, and it is urgent to implement clear application scenarios to achieve the leap from showing off skills to commercial use.

中国新闻网
5 min read
A supply and demand game sweeping the industrial chain
Developing·

A supply and demand game sweeping the industrial chain

Since the beginning of this year, cloud computing manufacturers and AI service providers such as Tencent Cloud, Alibaba Cloud, and Baidu Smart Cloud have intensively raised the prices of computing power products, with some services increasing by more than 30%. On the demand side, the explosive development of generative AI applications has led to a surge in the number of token calls. On the supply side, the supply side is constrained by factors such as tight production capacity of high-end AI chips and long construction cycles of intelligent computing centers. The imbalance between supply and demand has pushed up the price of computing power. Price increases are being transmitted upstream and downstream along the industrial chain, with upstream hardware manufacturers benefiting, midstream computing power leasing companies facing opportunities and pressures, downstream AI application companies facing increasing cost pressures, and industry differentiation may intensify. Experts suggest that companies should respond through technology cost reduction, resource scheduling optimization and strategic layout. The industry needs to move away from the pure hardware thinking and extend to algorithm optimization and scenario solutions.

中国新闻网
4 min read
More on this topic算力